知识卡片

普通读书笔记卡 · 9048.a · 生成模型:GPT-6

视频自监督依靠可重投影性

理解等级:中级

用深度网络估目标帧深度、位姿网络估两帧相对运动,再按已知内参将源图像重采样到目标帧,通过图像重建误差训练。无需人工深度标注,仍依赖标定、运动及光度假设;它是使用几何和时序信号的自监督,不是完全没有约束信息。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      As a camera moves in the world, the pixel values between two images captured at two different time instants are related by the 3D scene structure and the relative camera motion between the two frames. Therefore, we can formulate the problem as an image prediction problem. Let's say we have a function f1 that can estimate depth from a single image, and a function f2

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第43章《单图深度估计》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c043.xhtml) - 结论依据:摘录中关于视频自监督依靠可重投影性的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:As a camera moves in the world, the pixel values between two images captured at two different time instants are related by the 3D scene structure and the relative camera motion between the two frames. Therefore, we can formulate the problem as an image prediction problem. Let's say we have a function f1 that can estimate depth from a single image, and a function f2